Samsung Foundry confirms 2nm orders from hyperscalers as its first GAA chips enter production

Samsung Electronics confirmed during its third-quarter 2026 earnings presentation that its foundry division has secured manufacturing contracts with major cloud service providers (CSPs) and high-performance computing customers for AI. It is the first time the company has officially acknowledged that such customers have committed to its advanced 2nm process with Gate-All-Around (GAA) transistors. According to the TechTimes article, Sukchae Kang, executive vice president of the foundry business, said they had secured 2nm projects from major CSPs and AI and HPC customers in the third quarter, and that 2nm awards had risen significantly. Samsung also says 2nm project orders will double in 2026 compared with the previous year. The customers' names have not been disclosed.

The confirmation comes amid a record quarter. Samsung posted consolidated revenue of 195.3 trillion won (about $145.8 billion) and operating profit of 107.44 trillion won (about $80.1 billion) for the quarter ended September 30, 2026. According to the piece, it is the first tech company in the world to exceed 100 trillion won in operating profit in a single quarter. Operating profit grew nearly ninefold year on year (+783%, from 12.2 trillion won), and net profit reached 75.13 trillion won (+515%).

Those figures are almost entirely a memory story. The Device Solutions division (memory, System LSI and foundry) contributed 147.1 trillion won in revenue and 106.5 trillion won in operating profit, 99% of the total. Memory, driven by high-bandwidth memory (HBM) and server DRAM, set revenue and profit records, with roughly 50% quarter-on-quarter growth in HBM4 bit shipments according to analyst estimates. The rest of the business did not keep pace: the consumer electronics division (mobile, networks, TVs and appliances) posted an operating loss of about 800 billion won, of which about 600 billion won came from mobile and networks. The article attributes this to rising memory chip prices and competition from Chinese TV brands.

The article explains what Samsung's 2nm process (SF2) is. It uses a Gate-All-Around transistor architecture, a break from the FinFETs of the 5nm and 3nm generations: instead of the gate touching the channel on three sides, it surrounds it completely, improving electrostatic control and reducing the current leakage that drains batteries and generates heat. Samsung calls its nanosheet implementation MBCFET, in which the channel consists of several thin stacked silicon sheets surrounded by gate material. Samsung was the first to commercialize GAA at scale, with its 3nm node in 2022, about three years before TSMC made the same leap with its N2 in 2025.

The problem was yield. When GAA was introduced at 3nm, yield dropped because the architecture requires more manufacturing steps, more extreme ultraviolet lithography layers and more deposition and etch cycles, each a potential source of defects. According to the outlet, 2nm yield reached roughly 60% by the end of September 2026, still below the 70% that, according to the piece, Qualcomm and most designers require before committing to high-volume production, but well above the sub-20% level of late 2025. Samsung confirmed that 2nm GAA mass production began in the third quarter of 2026, which the article reads as a sign that at least some customers consider the effective yield acceptable.

The competitive context also matters. According to Counterpoint Research, cited in the article, TSMC holds 72% to 73% of the foundry market, and its most advanced 2nm capacity is committed to Nvidia, Apple, AMD and Qualcomm through at least 2028. In addition, TSMC's list prices for 2nm wafers are reportedly 50% higher than those of previous nodes. All of this is pushing hyperscalers and AI chip designers toward Samsung, described as the only commercially viable alternative to TSMC at 2nm.

On top of that comes the memory shortage. A day earlier, SK Hynix said on its own earnings call that its DRAM and HBM are virtually sold out through 2029, with a supply-demand imbalance expected to last until 2030. Samsung, for its part, warns in its earnings document that AI demand will far outstrip industry supply relative to 2026, and expects the chip market to remain supply-constrained through 2028.

A key driver of the foundry's recovery has been HBM4 memory. Samsung chose to manufacture HBM4 base dies on its own 4nm logic process, rather than the mature 12nm-class processes of previous generations. That consumes in-house capacity: according to the article, more than half of its main advanced node is devoted to this component for its memory division. As HBM4 demand grows, internal base die orders rise, filling the fabs, lowering unit fixed costs and improving results. Samsung attributes the foundry's improvement to growing HBM base die sales, higher sales on advanced nodes and continued 2nm orders from HPC customers, including major hyperscalers. The piece highlights Samsung's unusual position: it sells the memory, manufactures the logic inside that memory and is now winning orders to make those same customers' logic chips.

Despite all this, the foundry remained in operating losses in the third quarter, its fifth consecutive one. Samsung described this as expected: the one-off costs that weighed on previous quarters eased and fab utilization improved. For the fourth quarter, the company plans to begin mass production of new mobile products on its second-generation 2nm process (SF2P), which Samsung says offers 26% better power efficiency than SF2 thanks to design-technology co-optimization, and to ramp up full 4nm production for a large unidentified customer. It is targeting double-digit revenue growth, both quarter on quarter and year on year.

In the United States, Samsung is preparing to begin construction of a second fab in Taylor, Texas, in late 2026, with mass production slated for 2030. The first Taylor fab is set to begin mass production in 2027 and will initially serve Tesla, which signed a $16.5 billion chip supply agreement in July 2025 running through 2033. According to the article, it will be the first U.S. facility with commercial-scale 2nm, about two years ahead of TSMC's in Arizona.

The investment is substantial. Samsung reported cumulative capital expenditure of 56.9 trillion won in the chip division over the first three quarters of 2026, and expects the division's annual spending to reach 76.8 trillion won (+62%). Including Samsung Display, the total planned for 2026 is about 84.4 trillion won (about $63 billion).

Investor reaction was mixed: shares fell despite the record results. Shareholder returns comparable to SK Hynix's (about $75 billion) had been expected, but Samsung confirmed a quarterly dividend of 30 trillion won (4,601 won per share), more than twelve times that of the first two quarters. The full 2026 shareholder return program, estimated at between 90 and 110 trillion won, will be finalized at a board meeting in late January 2027. As for the foundry, profitability, which had been expected as early as the third quarter, now looks more likely in 2027 or 2028.

The article closes with a reflection on Samsung's structural advantage and risk. Unlike TSMC, a pure-play foundry that does not compete with its customers, Samsung is an integrated manufacturer that designs and sells its own memory and system-on-chip products. Hyperscalers that entrust it with their designs will have to trust that it won't use them for its own roadmap, and the safeguards Samsung builds in this regard will be as important as any yield or efficiency metric. Currency conversions in the piece are approximate (around 1,340 won per dollar). It should be noted that the content received cuts off before the end, so this summary is based only on the available portion of the article.